Sparse representation for ear biometrics

Abstract

In this paper we address for the first time, the problem of user identification using ear biometrics in the context of sparse representation. During the training session the compressed ear images are transformed to vectors to develop a dictionary matrix A [1]. The downsampled probe vector y is used to develop a linear, underdetermined system of equation y∈=∈Ax, x being unknown. The ill-posed system is regularized by utilizing the sparse nature of x and the inverse problem is solved through the l 1-norm minimization. Ideally the nonzero entries in the recovered vector x correspond to the class of the probe y. The developed system does not assume any preprocessing or normalization of the ear region. We did extensive experiments on the UND [2,3] and the FEUD [4] databases with session variability and incorporating different head rotations and lighting conditions. The proposed system is found to be robust under varying light and head rotations yielding a high recognition rate of the order of 98%. Moreover, in context of sparse representation a tuning parameter of the system is identified and is designated as an Operating Point (OP). The significance of the OP is highlighted by mathematical arguments and experimental verifications. © 2008 Springer Berlin Heidelberg.